Tom van der Woerdt created CASSANDRA-9728:
---------------------------------------------
Summary: Very bad read performance after 2.1.5->2.1.7 upgrade
Key: CASSANDRA-9728
URL: https://issues.apache.org/jira/browse/CASSANDRA-9728
Project: Cassandra
Issue Type: Bug
Reporter: Tom van der Woerdt
After upgrading from 2.1.5 to 2.1.7, read performance has significantly slowed
down for queries that use indexes. A query that used to take under a second to
execute now takes 8 seconds or longer.
The table I am noticing this on is a fairly large table (possibly pushing
Cassandra's limits a bit, with around 1TB per node) with a single index.
{noformat}
CREATE TABLE the_table (
id bigint PRIMARY KEY,
id_2 bigint,
field_1 ascii,
field_2 blob, -- "Wide" row in that it's on average 15KB
field_3 ascii,
field_4 ascii,
field_5 text,
field_6 ascii
) WITH compaction = {'class':
'org.apache.cassandra.db.compaction.LeveledCompactionStrategy'}
AND compression = {'chunk_length_kb': '1024', 'sstable_compression':
'org.apache.cassandra.io.compress.DeflateCompressor'};
CREATE INDEX the_table_id_2_idx ON the_table (id_2);
{noformat}
Compaction status, according to nodetool's cfstats :
index: SSTables in each level: [1714/4, 7, 0, 0, 0, 0, 0, 0, 0]
main: SSTables in each level: [79/4, 1, 105/100, 1049/1000, 172,
0, 0, 0, 0]
It's had a few days to catch up, but the indexes don't seem to be catching up
well. Do note that the performance issue I'm reporting was also experienced
when the main table had not yet caught up. Queries fetching by the primary key
have been equally fast before and after the compaction catch-up.
One thing I have noticed is that cfstats is reporting a *LOT* of reads on the
index, while there's no way I'm firing that many queries:
Local read count: 6670463
Local read latency: 4.134 ms
Based on the application's logs, I would say there shouldn't have been more
than 5000 queries against the index since Cassandra was started on this node.
What may also be relevant is that during the upgrade to 2.1.7, nodes that were
still on 2.1.5 were responding quickly (tested in cqlsh), while nodes that had
already upgraded were slow. Even when only one node was left to be upgraded,
that node was fast with responding.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
